DOF: Duplicate Detectable Opportunistic Forwardingin Duty-Cycled Wireless Sensor Networks

Daibo Liu1, Zhichao Cao2, Jiliang Wang2, Yuan He2, Mengshu Hou1, Yunhao Liu21School of Computer Science and Engineering, University of Electronic Science and Technology of China

2School of Software, TNLIST, Tsinghua University{dbliu, mshou}@uestc.edu.cn, {caozc, jiliang, he, yunhao}@greenorbs.com

Abstract—Opportunistic routing, offering relatively efficientand adaptive forwarding in low-duty-cycled sensor networks,generally allows multiple nodes to forward the same packetsimultaneously, especially in networks with intensive traffic.Uncoordinated transmissions often incur a number of duplicatepackets, which are further forwarded in the network, occupythe limited network resource, and hinder the packet deliveryperformance. Existing solutions to this issue, e.g. overhearingor coordination based approaches, either cannot scale up withthe system size, or suffers high control overhead. We presentDuplicate-Detectable Opportunistic Forwarding (DOF), a dupli-cate free opportunistic forwarding protocol for low-duty-cycledwireless sensor networks. DOF enables senders to obtain the infor-mation of all potential forwarders via a slotted acknowledgementscheme, so the data packets can be sent to the deterministicnext-hop forwarder. Based on light-weight coordination, DOFexplores the opportunities as many as possible and removesduplicate packets from the forwarding process. We implementDOF and evaluate its performance on an indoor test-bed with 20TelosB nodes. The experimental results show that DOF reducesthe average duplicate ratio by 90%, compared to state-of-the-art opportunistic protocols, and achieves 61.5% enhancement innetwork yield and 51.4% saving in energy consumption.

I. INTRODUCTION

Wireless sensor networks [1] [2] [3] [4] are usually dutycycled to prolong the network lifetime. A widely adopted low-duty-cycled media access mechanism is low power listening(LPL) [5]. Taking X-MAC [6] as a typical example of LPL,each node periodically wakes up and checks the receivedsignal strength to detect the potential traffic. If the channelis clear, it turns off the radio to sleep for a certain period.Note that the sleep schedule of different nodes is generallyunsynchronized. A sender probably has to spend much timewaiting for its corresponding forwarder to wake up. Duringthe waiting time, the sender continuously transmits the samedata packet (called preamble) till the preset timer expires or anacknowledgement is received. As a result, if the forwarder isdeterministic, the end-to-end delay is likely high. Obviously,sender energy is wasted on waiting for the forwarder. Theduty-cycled communication nature makes the deterministicforwarding schemes inefficient.

To shorten the waiting time, an intuitive idea is to takethe earliest forwarding opportunity instead of waiting forthe deterministic forwarder, like opportunistic routing [7].Temporally available links may be exploited to reduce thetransmission cost in wireless mesh networks. Landsiedel etal. propose ORW [8], an opportunistic forwarding protocol

for low-duty-cycled unsynchronized sensor networks. In ORW,any forwarder with certain routing progress can acknowledgethe preamble transmission in LPL. The first wake-up neighborthat successfully receives the packet is selected as the next-hopforwarder. Nevertheless, ORW cannot support high traffic loadapplications due to channel capacity degradation incurred bythe inherent duplicate problem.

Most duplicate packets are generated when several for-warders keep awake and receive the same data packet duringthe same period. In low-duty-cycled sensor networks, the hightraffic load will significantly increase the risk of producing du-plicates. Although several duplicate suppression mechanismsare proposed [7] [8] [9], the overhearing based approachesare not well adapted to the bursty traffic, especially in thelarge-scale networks with dynamic links. Moreover, accordingto MORE [10], the long coordination process diminishes thebenefits brought by opportunistic routing. The amount ofduplicate packets might increase exponentially along the multi-hop relay such that the network throughput is significantlydegraded.

In order to address the above issues, we propose DuplicateDetectable Opportunistic Forwarding (DOF). Instead of directdata transmission in LPL, a sender sends a probe and asksthe potential forwarders to acknowledge the probe respectivelyin different time slots. By utilizing the temporal diversity ofmultiple acknowledgements, the sender detects the quantityand differentiates the priority of all potential forwarders.The sender then forwards its data in the deterministic wayto avoid multiple forwarders hearing the same packets. Wedevelop methods to resolve possible collisions among multipleacknowledgments and exploit temporal long good links foropportunistic forwarding. With the light-weight mechanism tosuppress duplicates, DOF can adapt to various traffic loadsin duty-cycled sensor networks and enhances the systemperformance with respect to both network yield and energyefficiency.

The contributions of this work are as follows:

• Under the context of duty-cycled sensor networks, we ex-tend the opportunistic routing to fit the needs of various trafficloads. This work presents a more comprehensive solution tolow-duty-cycled opportunistic forwarding.

• We propose DOF, a practical duplicate-free opportunisticforwarding protocol by exploiting the temporal diversity ofthe acknowledgements. DOF minimizes the control overheadand improves the reliability of duplicate suppression. It can beeasily extended to opportunistic routing in other networks.978-1-4799-1270-4/13/$31.00 c©2013 IEEE

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

1

1 2 3 4 5 6 7 8 9 10 11Pro

babili

ty o

f m

ultip

le forw

ard

ers

Number of potential forwarders

λ = 1λ = 5

λ = 10λ = 20

(a) Poisson traffic model

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

1

1 2 3 4 5 6 7 8 9 10 11Pro

babili

ty o

f m

ultip

le forw

ard

ers

Number of potential forwarders

λ = 1λ = 5

λ = 10λ = 20

(b) Uniform traffic model

Fig. 1. The probability of multiple waking forwarders at one moment withdifferent number of potential forwarders. λ indicates the average traffic loadof each forwarder (packets per 10 seconds).

• We implement DOF and evaluate it on a testbed with20 TelosB nodes. In high traffic load settings, the evaluationresults show that DOF reduces the average duplicate ratioby 90%, compared to state-of-the-art protocols. Meanwhile,DOF achieves 61.5% enhancement in network yield and 51.4%saving in energy consumption.

The rest of the paper is organized as follows. Section IIpresents the motivation of this work. Section III introducesthe system design and analysis, followed by its implementationand evaluation in Sections IV and V, respectively. Section VIdiscusses the related work. Section VII concludes this paper.

II. MOTIVATION

In this section, we examine the performance degradationbrought by inherent duplicates of state-of-the-art opportunisticrouting in duty-cycled sensor networks. First, under differentprobabilistic models of traffic loads, e.g. poisson and uniformdistribution, we analyze the probability for multiple forwardersto wake up simultaneously under different network densitiesand protocol settings. Then, through testbed experiments ofORW, we show the relationship between the duplicates andsystem performance. Finally, we explain why the current du-plicate suppression mechanisms are inefficient in duty-cycledsensor networks.

A. Protocol Analysis

The forwarders that simultaneously keep awake may re-ceive the same data packet. The number of duplicates goes

0

0.5

1

0 0.5 1 1.5 2

CD

F

Duplicate ratio

20s10s

4s2s

0

0.5

1

0 0.2 0.4 0.6 0.8 1

CD

F

Radio duty cycle

0

0.5

1

0 0.2 0.4 0.6 0.8 1

CD

F

Packet reception ratio

Fig. 2. The CDF of the duplicate ratio, radio duty cycle, and packet receptionratio with different traffic loads on test-bed experiment, when the sleep intervalis 512ms.

TABLE I. THE DISTRIBUTION OF THE TRAFFIC LOAD CARRIED BYEACH NODE.

����������������� �����������

������������������������������������� ������ � �����

��� ����� ����� ������ �� ���� ������ ����� ����� ����� ����� ������ ����

up as the probability of multiple simultaneously waking for-warders increases. Due to the long preamble transmission ofLPL, data forwarding with LPL significantly prolongs thewaking time of the forwarders. Thus, the number of potentialforwarders and the traffic load can influence the probability ofmultiple forwarders being awake at one moment.

In the analysis, we assume each forwarder periodicallywakes up every 512ms. The forwarder stays awake for 20msafter it wakes up. The traffic model of a forwarder followseither poisson or uniform distribution. λ indicates the averagenumber of data packets passing each forwarder in 10s. Thetime of the preamble transmission of individual data forward-ing is calculated according to the traffic model. We simulate thedata forwarding process to calculate the probability of multiplewaking forwarders at one moment.

Figure 1 shows the probability distribution under poissonand uniform traffic model, respectively. In both scenarios, theprobability goes up with the increasing traffic load. For thesame traffic load, the probability increases as the number ofpotential forwarders increases. Thus, the duplicates tend toappear in the areas with bursty traffic or high node density.Specifically, in both Figures 1(a) and 1(b), when the averagetraffic load is 1 packet/second and there are 6 potentialforwarders, the probability of multiple simultaneously wakingforwarders is about 30% - 50%. This is much higher than thatin the low traffic load setting [8].

B. System Measurement

Based on the implementation of ORW in TinyOS, wefurther evaluate the influence of duplicates on a testbed with 25TelosB nodes. We set the radio power at 1 in TinyOS and thesleep interval is 512 ms. On the testbed, the minimum, medianand maximum number of available next-hop forwarders ofdifferent nodes are 1, 5 and 11, respectively. The average

�� ��

����

�������� ����

������ ����

(a) Topology

!

"�

"#

"$

"%

&� &� &�

�

�

&�

(b) Deterministic forwarding

!

"�

"#

"$

"%

&� &� �

&�

&�

&�

�

�

�

&�

(c) ORW

!

"�

"#

"$

"%

� � � �� � # $ &�

&�� �

�

�

�

�

�

�

�

�% � # $

(d) DOF

Fig. 3. Different from deterministic forwarding and ORW, in DOF the sender distinguishes the multiple waking forwarders by the temporal diversity of ACKs.Then it sends the data packet to an exclusive forwarder by adding in the ACK slot information.

length of routing paths is 2.08 hops. The maximum lengthis 7 hops. Each node generates data periodically. We selectfour different traffic loads with the inter packet interval (IPI)to be 2s, 4s, 10s and 20s, respectively. The actual distributionof the traffic load is shown in Table I.

According to the sequence number of the data packetsreceived by sink, we take the duplicate ratio, i.e., the numberof duplicates to the number of different packets received, asthe metric of duplicates. According to Figure 2, the duplicateratio is low when the traffic load is low. It increases quicklywith the increase of traffic load and the maximum duplicateratio reaches 200%. When the IPI is 2s, the duplicate ratio ofover 50% of nodes is higher than 100%.

We then take the radio duty cycle as the energy consump-tion indicator for a node. We can see a significant increase ofthe radio duty cycle when the duplicate ratio increases. Whenthe IPI is 2s, the radio duty cycle of over 40% of nodes ishigher than 60%. About half of the energy is wasted on thetransmission of duplicates. According to packet reception ratio(PRR) of Figure 2, we can see the PRR stays stable when thepacket interval is 4s, but it decreases quickly when the trafficload gets higher. The main cause of packet drops is forwardingqueue overflow, where the queue size is 10.

The experiments show that many duplicates indeed existwith the state-of-the-art low-duty-cycled opportunistic routingprotocols, especially when the traffic load is high. Moreover,the duplicates significantly degrade the system performanceand should be avoided.

C. Duplicate Suppression Mechanism

Most of existing duplicate suppression mechanisms arebased on overhearing. When a forwarder overhears a packet,which is identical to a pending packet in the forwarding queue,it deletes the packet from the queue. However, in currentsensor operating systems like TinyOS, the non-preemptive taskabstraction does not allow a node to interrupt on ongoingtransmission tasks. Moreover, the bursty traffic, especiallyin large-scale networks with dynamic links, further makes aforwarder hard to exactly overhear every packet relayed byothers.

To further reduce duplicates in the bursty traffic, packettransmissions are coordinated among different nodes in thenetwork. For example, ExOR [7] arranges the forwardingorder according to the routing progress and the quantity ofreceived data packets. According to More [10], however, thecoordination process introduces extra overhead. Moreover,

the coordination restricts concurrent transmissions and hencereduces the network yield.

We want to develop a forwarding approach, which candetect the simultaneous waking-up forwarders and inherentlyavoid the duplicates. Meanwhile, the forwarding approachkeeps the spatial diversity of the opportunistic routing as muchas possible.

III. SYSTEM DESIGN

DOF targets on developing a practical opportunistic for-warding scheme for various duty-cycled sensor network ap-plications. In this section, we discuss several issues: (1) theoverview of how DOF detects the potential forwarders by s-lotted acknowledgement (ACK), (2) the algorithm of ACK slotassignment and forwarding strategy, (3) the adaptive routingmetric. For simplicity we here illustrate the basic design ofDOF using X-MAC, a well adopted unsynchronized LPL MACas we mentioned above.

A. Overview of DOF

As Figure 3(a) shows, S sends packets to the intendeddestination R2. There are three potential relay nodes R1, R3and R4. The links are either reliable or bursty indicated as thesolid or dashed lines, respectively.

As Figure 3(b) shows, in traditional deterministic forward-ing, S continuously sends the data to the predetermined relaynode R4 till it wakes up. As Figure 3(c) shows, ORW takesthe early wake-up nodes (R1, R2 or R3) which receives thedata and provides routing progress as the next-hop forwarder.However, as Figure 3(c) shows, R1, R2 and R3 may receive thedata simultaneously. The duplicates then significantly degradethe system performance as introduced above.

DOF detects potential duplicates by using adaptive slottedACK when multiple forwarders are awake simultaneously. AsFigure 3(d) shows, instead of directly sending data, a sequenceof probes are first broadcast by S. The interval of two adjacentprobes is divided into multiple time slots. Each slot is longenough to receive an ACK. When R1, R2, and R3 receive oneprobe and any of them offers routing progress, each of themindependently selects a slot (2, 0, and 4) to send the ACK back.According to the slot information of the received ACKs (0 and4), S sends the data packet to a forwarder (R2) by adding inthe slot information (0).

To minimize the duplicates and keep the benefit of oppor-tunistic routing, the design of DOF faces several challenges:(1) Different forwarders should acknowledge the probe at

� �

� � � ��� � �� �� �� ��� �� �� �� �� ��� ��

������ ������ ������

��!

"���������#���$�

������%�&��%����'�

()*����� � � � � � �

Fig. 4. DOF splits all the ACK slots into 3 slightly overlapped priority zones.According to the routing progress, DOF randomly maps each forwarder intoa slot in different priority zones.

different slots. In addition, the routing progress of differentforwarders should be distinguished because the forwarder withmore routing progress should be used with a higher priority.(2) Although the communication overhead caused by probetransmissions for each data packet is little, it should be avoidedwhen the traffic load is high. (3) DOF may explore temporallyavailable links to forward data. However, the data ACK lossover these links may lead undesirable retransmissions due tothe bursty loss. So the short-term link performance should beconsidered.

B. ACK Slot Assignment

As we mentioned in the previous subsection, the tworequirements of ACK slot assignment are that multiple for-warders should be distributed into different slots and the sendershould infer the routing progress of different forwarders byACK slot distribution. As Figure 4 shows, the basic strategyis as follows: first, according to a hash function, the forwardermatches its routing progress Δ to a location Hsf on thepriority sequence. The priority sequence is like a ruler tomeasure the routing progress; then, we split all the ACK slotsinto multiple slightly overlapped zones, which are matchedto different segments of the priority sequence (e.g. zone 0→ {0− 9}); last, according to Hsf , we randomly assign oneslot in the selected zone.

There are six parameters in the calculation procedure, asshown in Table II. When forwarder f receives the probe sentby s, the routing progress is calculated by

Δsf = Ws −Wf (1)

Ws is carried in the probe and Wf is local routing information.If Δsf is larger than Δmax, we set it as Δmax. By Eq. (2),the routing progress Δsf is mapped to a location Hsf in thepriority sequence. A forwarder with a larger routing progressis mapped into the head area of the sequence.

Hsf = �(1− Δsf

Δmax) ∗N� (2)

f calculates in which ACK zone (zonef ) it should acknowl-edge the probe and the offset (δf ) in the segment of prioritysequence corresponding to zonef .

zonef = �Hsf · LN

� (3)

TABLE II. THE DESCRIPTION OF THE SYMBOLS IN THE ACKASSIGNMENT ALGORITHM.

!'�(� � &����������

)� �!��������%�+����$�,������-��!������$����.��

+�/ �!��+�/�+�+�������%�&��%������,�������!�&�*� �!�����������%�!��-��!��&����������#���$���� 0!����������+�����-�()*��������� �!����+�����-�&��������1�����-�()*�������"� �!����+�����-�()*������������$!�&��������1����

δf = Hsf − �zonef · NL� (4)

f randomly maps Hsf into the final ACK slot, slotf , as Eq. (5)shows.

slotf = zonef · �ML�+ �δf · L ·R

N�+ rand() (5)

where rand() is a random number between 0 and R. If slotfis larger than M , we make slotf equal to M .

Let’s take Figure 4 as an example. We assume L, Δmax, N ,M and R are 3, 5, 30, 10 and 4, respectively. If the forwarderprovides routing progress Δsf as 2.8, the location of prioritysequence, Hsf , equals to 13. Then, zonef and δf are 1 and 3.If we assume the random number rand() is 3, the slotf willbe the 7th slot.

Rather than assigning each forwarder a fixed ACK slot,our method is more flexible to utilize all temporarily availablelinks. Moreover, the parameters of our method are predeter-mined based on the local routing information so that there is noextra communication overhead. The computation complexityof the algorithm is low. However, this algorithm does notguarantee that multiple forwarders do not choose the sameACK slot. We show in practice this situation rarely happensin Section IV.

C. Forwarding Management

Note that a forwarder may serve multiple senders during ashort period. Each forwarder maintains a sender table, whichrecords the ACK slot information to trace the potential senders.Each entry of the sender table includes: the sender’s address,expected data sequence number (DSN), and the selected ACKslot.

When a probe is received, the forwarder first checks theattached routing metric Ws of the sender s. If the forwardercan provide routing progress (Δsf > 0), it selects an ACKslot slotf to acknowledge the sender. Then, if there is a recordof the same sender, the forwarder updates the correspondingrecord in the sender table. Otherwise, the forwarder adds a newentry into the table. Note that the DSN attaching in the receivedprobe copies that of the sender’s pending data packet. Uponthe acknowledged probe, the sender attaches the DSN and theselected ACK slot number as the virtual intended forwarderaddress. When the forwarder receives a data packet, it queriesthe sender table. If there is no matched entry, the forwarderdrops the packet and does nothing. Otherwise, it will take theresponsibility to forward the data packet.

Moreover, although the forwarder acknowledges the re-ceived probe, it still receives the duplicate of the same probe.

����

��

����

��

����

��

���� ���� ���� ��� ���� ��� ���� ���� ��� ��

���

��

��

���

����

���

����

�����

����

���

����

����

�

��� �!�"���#!�

����������������

��

�����

�����

�����

����

�����

�� �� �� �� � �� � �� �� �

$�

���

%&

���!!

��� ��

'�#(������!��)��!

��

�����

�����

�����

����

� � � � � �

�*�

��+�

� ���

!#�!

!���

����

�

'�#(���������,��)��!

����-��� ��*��.���#!����-��� ��*��.���#!

2�3� ����&��$������ 2�3�4����()*������������ 2$3�(,���%��.����&�$���������+�������$����� �

��

���

��

��

���

����

����

���

�� �� �� � �� � �� �� � �������������������������

���

���

��

��

$���

��#

!/#

���

��

0�#

-���

�� ��

)�+�

���

0�#���1����

)��� �#-��� ���

��

������

�����

������

�����

� � � � � �

$�-

����

����

��

'�#(���������,��)��!

2)����#�--��+����� ���$345!�#�--��+����� ���

��

����

���

���

����

��

�� �� � �� ��

0���

��� �

��!#

�!!�

�����

��

2� ���-��6� ��� ��*����!�

2.3�)��$��.��-�� � 2�3�)�+&����������5�����5�����������%�+������%����!+�� 2-3�"������������-�������������+�������

Fig. 5. The detail verification of DOF implementation. (a) relationship between the ACK slot time span and the prediction accuracy; (b) the situation of dataACK loss with multiple senders; (c) the average transmission count of the successfully delivered packet; (d) the clock drift as temperature changes; (e) duplicateratio of DOF slot assignment algorithm compared to the ideal assignment; (f) the ratio of tunnel transmissions over the total transmissions under different trafficloads.

The probe duplicate indicates the sender has not received theACK for the previous probe due to the asymmetric link or linkdynamics. If the forwarder receives the duplicate probe, it goesback to sleep to save energy.

On the other hand, the sender may receive multiple ACKsdistributed in different slots after sending a probe. Accordingto our ACK slot assignment algorithm, the forwarder corre-sponding to the earlier coming ACK provides relatively highrouting progress. Thus, the sender inserts the DSN and theminimum slot number of ACK received to the pending datapacket and sends it.

When the sender prepares to send a batch of packets,the intended forwarder will keep awake during the batchedsending. Besides the probes of the first packet, the probesof the rest packets are not needed. Thus, to save the extraoverhead of the probe transmission, the sender directly sendsthe rest packets with the connection (called Tunnel) found bythe probes of the first packet till either the loss of data ACKor there is no pending data packets.

When the pending data packet is acknowledged, the senderfinishes this transmission. However, because of the lossy linkor misalignment of the probe ACK slots, the sender may notreceive the data ACK from the intended forwarder. Hence,with a larger retransmission limit is inadvisable. Whether weshould keep retransmitting the data packet or send a probeagain to detect new forwarders is an important problem for theagility and efficiency of the protocol. According to [11] [12],the packet loss tends to be bursty over temporally availablelinks. We propose the Limited Retransmission Strategy (LRS)to address the data ACK loss. The basic idea is to estimatethe available period of those links and then adaptively boundthe number of retransmissions. The specific setting of LRS isshown in Section IV.

D. Low-Duty-Cycled Opportunistic Routing

In DOF, a packet is sent to one of the waking neighbors,which provides certain routing progress. As a result, therouting topology towards the sink is not fixed. A packet may beforwarded to the sink along different paths. Moreover, consid-ering the unsynchronized sleep schedule in LPL, DOF drivestwo requirements on routing. First, the routing metric shouldreflect the waiting time of the link layer transmissions. Second,each node should adaptively choose a set of forwarders fromall neighbors to determine the local routing metric.

Considering the two requirements above, EDC (expectedduty cycle), which is introduced by ORW [8], acts well on thewhole. Hence, we adopt the concept of EDC as the routingmetric. Our method of duplicate detection can be easily builton other routing metrics as well, such as end-to-end delay orETX[13].

IV. IMPLEMENTATION

We implement DOF on TelosB nodes in TinyOS 2.1.1. TheRAM and ROM consumption of the program are 6268 bytesand 39714 bytes, respectively. Next, several implementationissues are carefully discussed.

A. ACK Slot Settings

For the forwarder, when the radio (CC2420) has receiveda packet, it will generate an interrupt (FIFOP) to triggerthe handler function. Meanwhile, for the sender, it will alsogenerate an interrupt (falling-edge SFD) when the transmissionhas finished. We neglect the propagation delay so that theFIFOP interrupts of the probe on different forwarders happensimultaneously. Thus we take the simultaneous interrupts asthe beginning of ACK response for the sender and forwarders.

When the FIFOP interrupt is generated, the forwardercalculates its ACK slot Kf and sends the ACK after KfTslot

time, where Tslot is the slot time span. Upon receiving theACKs from different forwarders, the sender needs to determinethe slot number for different forwarders. Since the ACK slotcalculation and ACK reception take certain time, there is a shiftbetween the time the probe is sent and the time the first ACKis received. We denote the shift between the earliest receivedACK (i.e. the ACK sent in the first slot) and the falling-edgeSFD interrupt as Tbase. Tbase is close to a constant time. Inour implementation, the measured Tbase is about 2.3ms. Thus,according the interval Tr from the falling-edge SFD interruptto the ACK received, the sender calculates the slot number Ks

as:Ks = �(Tr − Tbase)/Tslot� (6)

where Ks should equal to Kf . If the maximum ACK waitingduration is Tmax, the maximum number slotmax of the ACKslots is calculated as:

slotmax = �(Tmax − Tbase)/Tslot� (7)

In practice, the clock drift and the variance of the softwareexecution time will incur the mismatch between Kf and Ks.Assume the clock drift between the forwarder and sender isσ, a mismatch only occurs when

σTr ≥ Tslot (8)

Thus, increasing the slot time span reduces the probability ofmismatch, while it limits the number of available ACK slots.In practice the less available ACK slots might increase theprobability that multiple receivers choose the same ACK slot.

We conduct experiments to show the prediction accuracyand the average time variance with different slot time spans.In the experiments, two senders transmit packets to the samereceiver in the office environment. The results are shown inFigure 5(a). We can see the predication accuracy is close to100%, when the slot time span is larger than 0.1ms. Theaverage variance is relatively stable, about 0.5 jiffy (1 jiffy= 1/32 ms). Considering the more complicate environment inpractice, the slot time span is conservatively set as 0.2ms andthe total number of ACK slots is 10.

Moreover, we measure the clock drift under differenttemperature between a pair of nodes. As Figure 5(d) shows,the clock drift goes up when temperature rises from hour 11 tohour 13. The clock drift is less stable when the temperature ishigher than 35 degree centigrade. The maximum clock drift isabout 140ms per minute (about 0.0023ms per 1ms). Under theabove settings, the maximum ACK waiting duration is about4.3ms. Thus, the maximum variance incurred by the clock driftis about 0.009ms, which is far less than 0.2ms.

B. Data ACK Loss and Retransmission

As Section III-C mentioned, the data ACK may be lostdue to the mismatch of the probe ACK slot between senderand forwarder or link dynamics. Figure 5(b) shows the dataACK loss rate when multiple senders send packets periodicallyto the same receiver. The data ACK loss ratio is no morethan 3% with different number of senders. In DOF, the senderopportunistically utilizes the temporally available links, forwhich the probe ACKs have been successfully received. Thus,the data ACK loss is rare in the experiments. The probe ACK

TABLE III. SYSTEM PARAMETERS.

�������� + ��

� ��

� �

" �

* ��

! ������������ ���+��

�"!��,���,� �� �

��� �

loss will reduce the available opportunities, but not degradethe reliability when the firm links exist.

Due to the possible bursty loss, we propose the limitedretransmission strategy (LRS) to bound the number of dataretransmissions. To determine the maximum retransmissioncount in practical networks, we make multiple senders sendpackets to the same receiver. For each packet, the maximumnumber of retransmission is initially set to 10. The packetwill be dropped when the retransmission is larger than 10.We measure the average transmission count of the successfuldata transmissions of all senders. As shown in Figure 5(c),the average transmission count of a successful transmission issmaller than 2. Thus, we set the transmission threshold as 2.When the sender does not hear the ACK, it will retransmit thedata packet once. If the retransmission also fails, the senderwill broadcast the probe to find the available forwarder again.

C. Tunnel Transmission

The probe is utilized to detect the potential receivers. AsSection III-C mentioned, when there are several packets inforwarding queue, the sender will take the tunnel transmissionto save the energy consumption on the probe transmission.Figure 5(f) shows the ratio between the number of tunneltransmissions and total data transmissions for different trafficloads on testbed experiment with 20 TelosB nodes. We couldsee for high traffic loads the portion of tunnel transmissionratio is high. Especially, over 50% of transmissions are tunneltransmissions when the inter packet interval is 1s.

D. Slot Assignment

As mentioned in Section III-B , the slot assignmentalgorithm of DOF may map different forwarders into thesame ACK slot. In this situation, it is possible that multipleforwarders receive the same packet so that duplications occur.We compare our algorithm with the ideal slot assignment, inwhich we manually assign an unique ACK slot for each of theforwarders. Figure 5(e) shows that, in practice, the duplicateratio of our slot assignment algorithm is just a little higherthan the ideal method. The duplicate in the ideal assignmentalgorithm is induced by the data ACK loss. In such a case, thesender will transmit the data packet again to a new forwarder,while the previous forwarder actually has received the packet.

In summary, we show the details of the implementationsettings in Table III.

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1 2 4 8 16

PR

R

IPI (s)

DOF

ORW

CTP-AMAC

CTP-XMAC

(a) Packet Reception Ratio

0

0.2

0.4

0.6

0.8

1

1 2 4 8 16

Duty

cycle

IPI (s)

DOF

ORW

CTP-AMAC

CTP-XMAC

(b) Radio Duty Cycle

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1 2 4 8 16

Duplic

ate

ratio

IPI (s)

DOF

ORW

CTP-AMAC

CTP-XMAC

(c) Duplicate Ratio

0

4

8

12

16

20

1 2 4 8 16

Avera

ge p

ream

ble

count

IPI (s)

DOF

ORW

CTP-XMAC

(d) Average Preamble Count

Fig. 6. The overall comparison of different system performance metrics between DOF and the other forwarding protocols.

V. EVALUATION

In this section, we evaluate DOF through various testbedexperiments. We compare the network yield, energy consump-tion and duplicate ratio of DOF with two unsynchronizedlow-duty-cycled forwarding protocols such as ORW and CTP-XMAC. We also compare the performance of DOF with CTP-AMAC [14], which is the-state-of-art synchronized receiver-initiated low-duty-cycled protocol. In addition, consideringthe system stability under network churns, we compare DOFwith L2 [15], which is proposed to optimize the energyefficiency by incorporating with the synchronized rendezvous,link burstiness and dynamic forwarding.

We use the packet reception ratio (PRR) as the indicatorof the network yield. It also indicates the network throughputcombining with the inter packet interval. In our implementa-tion, the on-board sensor and flash memory is rarely used andthus the radio consumes most of the energy [16]. The energyconsumption is measured by the radio duty cycle. Moreover,we use the average preamble count to approximate the delay.In CTP-XMAC and ORW, it is the average number of the datatransmissions. In DOF, it is the sum of the probes and datatransmissions on average. Normally, the smaller the averagepreamble count is, the less the delay is.

A. Evaluation Setup

We evaluate the performance of different forwarding proto-cols on an indoor testbed with 20 TelosB sensor nodes. We set

the transmission power of CC2420 as 1 to ensure multi-hopcommunication (maximum hop is 3).

Each node generates packets with a fixed IPI. We vary thetraffic load by setting different IPIs, such as 1s, 2s, 4s, 8s, and16s. The packet length is 80 bytes. Thus, the sleep interval ofother protocols is set to 512ms, except AMAC which is set tothe default setting 128ms. For each traffic load, the experimentslast for at least 30 minutes and are repeated three times. Theexperiments are often conducted during the night to mitigatethe influence of the human behavior.

B. Network Yield

Figure 6(a) shows the experiment results of the averagePRR for different forwarding protocols with different trafficloads. We can see that when the IPI is no less than 8s, thepacket loss of all forwarding protocols is small. However, withthe decreasing of IPI, the PRR of ORW, CTP-AMAC, andCTP-XMAC sharply decreases from 95% to less than 50%. Incontrast, the PRR of DOF is still higher than 90% and 70%when the IPI is 2s and 1s, respectively. When the IPI is 1sand 2s, the network yield of DOF is about 46.5% and 61.5%higher than the best of ORW, CTP-AMAC, and CTP-XMAC.

The significant decreasing of PRR in a high traffic loadis due to the inefficient channel utilization. In CTP-XMAC,each sender will occupy the channel for a long time till theintended receiver wakes up. As shown in Figure 6(d), theaverage preamble count of CTP-XMAC is much larger thanboth ORW and DOF. Although ORW has the smallest average

������

������

������

������

�

$�-

����

����

��

378$34

9:�%

��

����

��

����

��

$�-

����

����

��

378$34

9:�%

��

����

����

����

���

����

$�-

����

����

��

378$34

9:�%

��

����

���

���

����

��

$�

����

���

��

����

���

���

����

��

$�

����

���

�����������������������������������

��

$�

����

���

��

�

��

���

��

���

� � � � � � �� �� �� �� � �� � �� �� �

�*�

��+�

�-��

�#(�

����

��

'�)��2$

��

�

��

���

��

���

� � � � � � �� �� �� �� � �� � �� �� �

�*��

�+��

-���

#(�

����

��

'�)��2$

��

�

��

���

��

���

� � � � � � �� �� �� �� � �� � �� �� �

�*��

�+��

-���

#(�

����

��

'�)��2$

� 2�3� � 6"67��� 2�3� � 6"67��� 2$3� � 6"67���

Fig. 7. Per node comparison of different system performance metrics between DOF and the other forwarding protocols in high traffic loads.

preamble count, the duplicate ratio is much higher than othersas shown in Figure 6(c). The duplicate degrades the channelutilization, as explained in Section II. In AMAC, instead ofthe continuous data packet transmission, the sender waits forthe probe from the receiver when it wakes up. Upon receivinga probe, the sender sends the pending packet to the receiverimmediately. When multiple senders have packets for the samereceiver, packet collisions occur and data retransmissions willsignificantly reduce the channel utilization. Hence, DOF ismore adaptive for various traffic loads and thus the networkyield of DOF is better than others.

C. Energy Consumption

As Figure 6(b) shows, the average duty cycle of DOFis the smallest among various traffic loads, except for thecase of CTP-AMAC when the traffic load is low (IPI = 16s).AMAC is high energy efficient in low traffic load because ituses one-hop synchronization. The advantages of transmissionsolicitation and synchronization in AMAC transform to theweakness as the traffic load increases. This is because thesolicitation from the receiver to effectively synchronize packettransmissions results in concentration in packet transmission,and thus contention and collision. DOF saves at least 21.4%and 51.4% of energy compared with others when IPI is 1sand 2s, respectively. The CTP-XMAC has the lowest energyefficiency. ORW performs badly when IPI is less than 4s. Theenergy consumption of CTP-AMAC increases slowly when IPIis less than 8s.

In a high traffic load, the sharp increasing of energyconsumption of CTP-XMAC and ORW is due to the degra-dation of the channel utilization mentioned above. However,the energy consumption in CTP-AMAC increases slowly. Weguess the reason is the synchronized sleep schedule of AMAC.

The histograms in the middle of Figure 7 clearly show theduty cycle of each node for different protocols with differenttraffic loads. When the IPI is 1s and 2s, on most of nodes,the duty cycle of DOF is much better than ORW and CTP-XMAC. ORW consumes almost the same energy as CTP-XMAC. When IPI is 4s, on most of nodes, the duty cycle ofORW is close to that of DOF, and both of them are better thanCTP-XMAC. The results verify that although ORW performswell in a low traffic load network, DOF can keep the energyconsumption low for various traffic loads.

DOF utilizes the probe to detect potential forwarders so thatit induces a low communication overhead. The experimentsshow that in the high traffic load, the energy efficiency broughtby the probe is much larger than the overhead. In a low trafficload, the overhead of the probe transmission is also limited.

D. Duplicate Ratio

In Figure 6(c), compared with ORW, DOF significantlyreduces duplicate ratio when IPI is less than 4s. The averageduplicate ratio of ORW is about 85% when IPI is 1s. This isabout 10 times larger than DOF. The duplicate ratio of DOF iscomparable with the deterministic routing under various trafficloads. As the histograms on the top of Figure 7 show, on mostof the nodes, the duplicate ratio of DOF and CTP-XMAC aremuch less than ORW. When IPI is 1s, the highest duplicateratio of ORW exceeds 300% (e.g., node 19). Compared toORW’s high duplicate ratio, DOF always keeps the duplicateratio low, which is close to the duplicate ratio of CTP-XMACin different traffic loads. The results of duplicate ratio verifythe efficiency of the ACK slot assignment algorithm.

E. Delay

As Figure 6(d) shows, the average preamble count of CTP-XMAC do not change significantly with the increasing of

network traffic load. The average preamble count of DOF,which is comparable with ORW, is 3 times less than CTP-XMAC. The average preamble count increases when the trafficload increases. The histograms on the bottom of Figure 7 showthe average preamble count of each node in different trafficloads. We can see CTP-XMAC always has the largest delayon every node. Due to channel degradation, the maximumaverage preamble count reaches 8 and 12 when IPI is 2s and1s, respectively. DOF has less delay than ORW in the scenarioswhen IPI is no more than 2s, because the waking forwarderssimultaneously send ACKs resulting in ACK collisions inORW protocol. Then, the sender will broadcast the data packetagain. In a low traffic load, the overhead brought by probetransmission makes the average preamble count of DOF alittle greater than that of ORW in single-hop propagation.Considering the preferential use of the links with high routingprogress, we believe that DOF can reduce hop count comparedto ORW.

F. Impact of Network Churn

The node reboot, node replacement or link dynamic willincur network churn, which may further lead to the degradationof system performance. We evaluate the impact of networkchurn on DOF and L2. L2 optimizes the energy utility byincorporating the synchronized rendezvous, link burstiness anddynamic forwarding. In the implementation of L2, each nodewill take 10 minutes to synchronize the rendezvous withneighbors by routing beacons before it begins to generatedata. When the synchronization is stable, the frequency of therouting beacon is reduced to one per several minutes.

We set the IPI as 4s and do the experiments for one hourin the daytime with the influence of the human behavior andWiFi. Moreover, for DOF, we randomly remove and add nodes.For L2, we randomly reboot nodes. The top figure of Figure 8shows the number of nodes sending packets.

The middle figure of Figure 8 illustrates the variance ofthe PRR. We can see that there is almost no influence onDOF since there is no need of any extra control message inDOF. However, we can see the PRR of L2 decreases evenwithout churn. The reason is that the a high traffic load willlead to time error accumulation in TinyOS system so thatthe synchronization accuracy will decrease quickly. Withoutenough routing beacons to re-synchronization, routing loops ordata retransmissions will significantly degrade the PRR. Whenthere are network churns, the PRR of L2 tends to be moredynamic.

The bottom figure of Figure 8 illustrates the variance ofthe radio duty cycle along the time. We can see that theenergy consumption of DOF without churn is relatively stable.When there are network churns, since the reduced data amountbrings low channel contention, the energy consumption couldbe reduced as shown around the 15th time unit. However,if nodes around the sink are removed, the path length willincrease. Thus the churn might also increase the energyconsumption as shown around the 22th time unit. The window-based transmission of L2 makes the energy consumption of asingle transmission is bounded, even with packet loss. We canalso see the energy consumption of L2 last increasing slowly.

��

��

���

���

���

'�#

(����

����)

�!

'���1���%1���

��

����

���

���

����

��

;77

<�=�����1���<�=��1���

$34=�����1���$34=���1���

��

����

����

����

�� �� ��� ��� ��� ��� ���

$�

����

���

� � � � � � � 0�+������

Fig. 8. Impact of churn on DOF and L2 in the network with IPI=4s andremoving or adding several nodes. The x-axis indicates the time units. Eachunit is corresponding to 2 minutes.

Although L2 has less energy consumption, it needs extracontrol overhead, i.e. routing beacon, to deal with networkchurns and thus the energy consumption increases. In contrast,DOF explores the temporally available links by probe, but doesnot need up-to-date link state maintenance. So DOF is moreadaptive to practical large scale network deployments.

G. Discussion and Limitations

DOF uses software acknowledgement in the implemen-tation. However, in current CC2420 radio stack in TinyOSsystem, we found the software ACK is vulnerable when thetraffic load is high. The limitation is due to the slow bufferswapping between MCU and CC2420. One consequence isthat a new packet may arrive when the sender is waiting forthe ACK. The processing time of the received packet mightaffect the accuracy of the ACK slot calculation. The otherconsequence is increasing the collision probability betweenACK and data packets.

For DOF in a high traffic load, the transmission of probe,data packet and ACK are mixed, which might lead to ACKloss or ACK slot prediction error. This explains why the PRRin Figure 5(a) is lower than 80% when IPI is 1s. The averagepreamble count when IPI is 1s is higher than that when IPIis 2s. We believe that DOF could work better with more fine-grained timing control in the radio stack.

VI. RELATED WORK

In this section, we discuss related work on opportunisticand dynamic forwarding mechanisms. Moreover, we illustratethe advantage of our adaptive duplicate suppression schemesin unsynchronized duty-cycled WSN.

ExOR [7] develops a complete opportunistic routing forwireless network. ExOR assigns each receiver to furthertransmit in distinct time slot, the receiver overhears others’transmissions to avoid the duplicates. MORE [10] targets onthe inefficient coordination process of ExOR and proposesa coding approach to eliminate the overhead. Rather thannetwork coding, DOF takes a light weight method to mitigatethe overhead for WSN.

BRE [12] develops the overhearing scheme on CTP [17]to capture the temporally good links. The sender changes thenext-hop receiver when the opportunity appears to reduce thetransmission count. However, BRE does not address the dutycycle issue, in which the waiting time dominates the energyefficiency.

In DSF [18], each node knows when the schedule ofneighbor nodes by synchronization. DSF dynamically selectsmultiple next-hop forwarders based on the sleep schedules androuting metrics of the neighbors. L2 [15] further notices thelink bursty to optimize the energy consumption on each packetand improve the network yield. However, DSF and L2 needextra control overhead to stabilize the forwarding schedule,which is vulnerable to dynamic links and network churn.

ORW [8] implements the opportunistic routing for un-synchronized low-duty-cycled WSN, but shows the limitedperformance for high traffic load applications. DOF extendsthis work to more general purpose WSN applications. CMAC[9] includes the slotted acknowledgements, but CMAC still de-termines the unique forwarder by overhearing other’s acknowl-edgments. In DOF, the sender distinguishes the forwarders,and then considers the link quality to arrange the forwardingschedule.

There are also some theoretical works focusing on oppor-tunistic routing [19] [20] and dynamic forwarding [21] [22] forwireless sensor networks. Although the models and simulationshow the efficiency of the opportunistic routing, they neglectthe practical issues addressed by DOF.

VII. CONCLUSION

Developing an adaptive and efficient forwarding protocol isurgent for duty-cycled wireless sensor network. In this paper,we propose DOF, a duplicate-detectable unsynchronized low-power opportunistic forwarding which is adaptive to varioustraffic loads. Based on the slotted acknowledgement, DOFmainly solves the channel degradation problem incurred bythe large amount of duplicates in traditional opportunisticforwarding and retains the benefits of the opportunistic routingas much as possible. The testbed experiments show DOFis more efficient and reliable than state-of-the-art low-duty-cycled forwarding protocols.

ACKNOWLEDGMENT

This study is supported in part by NSFC under Grant61073177, National Basic Research Program (973) under

Grant No. 2014CB347800, NSFC under Grant 61202359, NS-FC Major Program 61190110, NSFC under Grant 61170213,and the Fundamental Research Funds for the Central Univer-sities under grant ZYGX2011J062.

REFERENCES

[1] X. Mao, X. Miao, Y. He, X.-Y. Li, and Y. Liu, “Citysee: Urban CO2

monitoring with sensors,” in Infocom. IEEE, 2012, pp. 1611–1619.[2] L. Mo, Y. He, Y. Liu, J. Zhao, S.-J. Tang, X.-Y. Li, and G. Dai, “Canopy

closure estimates with greenorbs: sustainable sensing in the forest,” inSensys. ACM, 2009, pp. 99–112.

[3] M. Ceriotti, M. Corra, L. D’Orazio, R. Doriguzzi, D. Facchin, S. Guna,G. P. Jesi, R. L. Cigno, L. Mottola, A. L. Murphy et al., “Is there lightat the ends of the tunnel? wireless sensor networks for adaptive lightingin road tunnels,” in IPSN. IEEE, 2011, pp. 187–198.

[4] X. Wu, M. Liu, and Y. Wu, “In-situ soil moisture sensing: Optimalsensor placement and field estimation,” TOSN, vol. 8, no. 4, p. 33,2012.

[5] J. Polastre, J. Hill, and D. Culler, “Versatile low power media accessfor wireless sensor networks,” in Sensys. ACM, 2004, pp. 95–107.

[6] M. Buettner, G. V. Yee, E. Anderson, and R. Han, “X-mac: a shortpreamble mac protocol for duty-cycled wireless sensor networks,” inSensys. ACM, 2006, pp. 307–320.

[7] S. Biswas and R. Morris, “Exor: opportunistic multi-hop routing forwireless networks,” in ACM SIGCOMM Computer CommunicationReview, vol. 35, no. 4. ACM, 2005, pp. 133–144.

[8] O. Landsiedel, E. Ghadimi, S. Duquennoy, and M. Johansson, “Lowpower, low delay: opportunistic routing meets duty cycling,” in IPSN.ACM, 2012, pp. 185–196.

[9] S. Liu, K.-W. Fan, and P. Sinha, “Cmac: an energy-efficient maclayer protocol using convergent packet forwarding for wireless sensornetworks,” TOSN, vol. 5, no. 4, p. 29, 2009.

[10] C. Szymon, J. Michael, K. Sachin, and K. Dina, “More: network codingapproach to opportunistic routing,” MIT-CSAIL-TR-2006-049, 2006.

[11] K. Srinivasan, M. A. Kazandjieva, S. Agarwal, and P. Levis, “The β-factor: measuring wireless link burstiness,” in Sensys. ACM, 2008, pp.29–42.

[12] M. H. Alizai, O. Landsiedel, J. A. B. Link, S. Gotz, and K. Wehrle,“Bursty traffic over bursty links,” in Sensys. ACM, 2009, pp. 71–84.

[13] D. S. De Couto, D. Aguayo, J. Bicket, and R. Morris, “A high-throughput path metric for multi-hop wireless routing,” WirelessNetworks, vol. 11, no. 4, pp. 419–434, 2005.

[14] P. Dutta, S. Dawson-Haggerty, Y. Chen, C.-J. M. Liang, and A. Terzis,“Design and evaluation of a versatile and efficient receiver-initiated linklayer for low-power wireless,” in Sensys. ACM, 2010, pp. 1–14.

[15] Z. Cao, Y. He, and Y. Liu, “L2: Lazy forwarding in low duty cyclewireless sensor networks,” in Infocom. IEEE, 2012, pp. 1323–1331.

[16] R. Fonseca, P. Dutta, P. Levis, and I. Stoica, “Quanto: Tracking energyin networked embedded systems,” in OSDI. USENIX Association,2008, pp. 323–338.

[17] O. Gnawali, R. Fonseca, K. Jamieson, D. Moss, and P. Levis, “Collec-tion tree protocol,” in Sensys. ACM, 2009, pp. 1–14.

[18] Y. Gu and T. He, “Data forwarding in extremely low duty-cycle sensornetworks with unreliable communication links,” in Sensys. ACM, 2007,pp. 321–334.

[19] X. Mao, X.-Y. Li, W.-Z. Song, P. Xu, and K. Moaveni-Nejad, “Energyefficient opportunistic routing in wireless networks,” in MSWiM. ACM,2009, pp. 253–260.

[20] G. Schaefer, F. Ingelrest, and M. Vetterli, “Potentials of opportunisticrouting in energy-constrained wireless sensor networks,” in WirelessSensor Networks. Springer, 2009, pp. 118–133.

[21] J. Kim, X. Lin, and N. Shroff, “Optimal anycast technique for delay-sensitive energy-constrained asynchronous sensor networks,” in Info-com, 2009, pp. 612–620.

[22] J. Kim, X. Lin, N. B. Shroff, and P. Sinha, “On maximizing the lifetimeof delay-sensitive wireless sensor networks with anycast,” in Infocom.IEEE, 2008, pp. 807–815.